lee/OptiScaler
1
0
Fork 0
mirror of https://github.com/optiscaler/OptiScaler.git synced 2026-09-23 05:46:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixes for SL inputs, incorrect resource size, invalid cmdlist, doubled slSetConstants, remove unnecessary hooks, fix sl dll override hooking

Browse Source
This commit is contained in:
FakeMichau
2025-08-12 17:20:48 +02:00
parent b8061e0e10
commit 889ceee24b
9 changed files with 117 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files
OptiScaler/framegen/Streamline_Inputs_Dx12.cpp
+77 -29
View File
@@ -21,10 +21,13 @@ bool Sl_Inputs_Dx12::setConstants(const sl::Constants& values, uint32_t frameId)
LOG_DEBUG("Setting consts for streamline index: {}", index);
// Streamline logs already log this
// Streamline already log this error, keep using the previous data
static uint32_t lastFrameId = UINT32_MAX;
setConstantsSameFrameId = lastFrameId == frameId;
lastFrameId = frameId;
if (lastFrameId == frameId)
{
lastFrameId = frameId;
return false;
}
data = sl::Constants {};
@@ -147,8 +150,7 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
return true;
}
// TODO: detect if there are multiple calls to hkslSetConstants with the same ID, DRG crashing, sl sees an error
if (dispatched && !setConstantsSameFrameId)
if (dispatched)
{
fgOutput->StartNewFrame();
@@ -167,23 +169,32 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
hudlessSent = true;
// ResTrack_Dx12::SetHudlessCmdList(cmdBuffer);
auto hudlessResource = (ID3D12Resource*) tag.resource->native;
auto validity =
(tag.lifecycle != sl::eOnlyValidNow) ? FG_ResourceValidity::UntilPresent : FG_ResourceValidity::ValidNow;
// We need to make sure we have a copy of hudless that WE can use
if (cmdBuffer != nullptr && validity == FG_ResourceValidity::ValidNow)
validity = FG_ResourceValidity::ValidButMakeCopy;
// TODO: bugged and copy crashes
// if (cmdBuffer != nullptr && validity == FG_ResourceValidity::ValidNow)
// validity = FG_ResourceValidity::ValidButMakeCopy;
fgOutput->SetResource(FG_ResourceType::HudlessColor, cmdBuffer, hudlessResource, tag.extent.width,
tag.extent.height, (D3D12_RESOURCE_STATES) tag.resource->state, validity);
auto width = tag.extent.width;
auto height = tag.extent.height;
if (!tag.extent)
{
const auto desc = hudlessResource->GetDesc();
width = desc.Width;
height = desc.Height;
}
fgOutput->SetResource(FG_ResourceType::HudlessColor, cmdBuffer, hudlessResource, width, height,
(D3D12_RESOURCE_STATES) tag.resource->state, validity);
// Assume hudless is the size used for interpolation
interpolationWidth = tag.extent.width;
interpolationHeight = tag.extent.height;
interpolationWidth = width;
interpolationHeight = height;
auto static lastFormat = DXGI_FORMAT_UNKNOWN;
auto format = hudlessResource->GetDesc().Format;
@@ -206,9 +217,20 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
auto depthResource = (ID3D12Resource*) tag.resource->native;
auto width = tag.extent.width;
auto height = tag.extent.height;
if (!tag.extent)
{
const auto desc = depthResource->GetDesc();
width = desc.Width;
height = desc.Height;
}
const auto validity =
(tag.lifecycle != sl::eOnlyValidNow) ? FG_ResourceValidity::UntilPresent : FG_ResourceValidity::ValidNow;
fgOutput->SetResource(FG_ResourceType::Depth, cmdBuffer, depthResource, tag.extent.width, tag.extent.height,
fgOutput->SetResource(FG_ResourceType::Depth, cmdBuffer, depthResource, width, height,
(D3D12_RESOURCE_STATES) tag.resource->state, validity);
}
else if (tag.type == sl::kBufferTypeMotionVectors)
@@ -217,16 +239,25 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
mvsSent = true;
mvsWidth = tag.extent.width;
mvsHeight = tag.extent.height;
// ResTrack_Dx12::SetMVsCmdList(cmdBuffer);
auto mvResource = (ID3D12Resource*) tag.resource->native;
auto width = tag.extent.width;
auto height = tag.extent.height;
if (!tag.extent)
{
const auto desc = mvResource->GetDesc();
width = desc.Width;
height = desc.Height;
}
mvsWidth = width;
mvsHeight = height;
const auto validity =
(tag.lifecycle != sl::eOnlyValidNow) ? FG_ResourceValidity::UntilPresent : FG_ResourceValidity::ValidNow;
fgOutput->SetResource(FG_ResourceType::Velocity, cmdBuffer, mvResource, tag.extent.width, tag.extent.height,
fgOutput->SetResource(FG_ResourceType::Velocity, cmdBuffer, mvResource, width, height,
(D3D12_RESOURCE_STATES) tag.resource->state, validity);
}
else if (tag.type == sl::kBufferTypeUIColorAndAlpha)
@@ -237,13 +268,22 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
if (uiRequired)
{
// ResTrack_Dx12::SetUICmdList(cmdBuffer);
auto uiResource = (ID3D12Resource*) tag.resource->native;
auto width = tag.extent.width;
auto height = tag.extent.height;
if (!tag.extent)
{
const auto desc = uiResource->GetDesc();
width = desc.Width;
height = desc.Height;
}
const auto validity = (tag.lifecycle != sl::eOnlyValidNow) ? FG_ResourceValidity::UntilPresent
: FG_ResourceValidity::ValidNow;
fgOutput->SetResource(FG_ResourceType::UIColor, cmdBuffer, uiResource, tag.extent.width, tag.extent.height,
fgOutput->SetResource(FG_ResourceType::UIColor, cmdBuffer, uiResource, width, height,
(D3D12_RESOURCE_STATES) tag.resource->state, validity);
}
}
@@ -253,14 +293,23 @@ bool Sl_Inputs_Dx12::reportResource(const sl::ResourceTag& tag, ID3D12GraphicsCo
distortionFieldSent = true;
// ResTrack_Dx12::SetDistortionFieldCmdList(cmdBuffer);
auto distortionFieldResource = (ID3D12Resource*) tag.resource->native;
auto width = tag.extent.width;
auto height = tag.extent.height;
if (!tag.extent)
{
const auto desc = distortionFieldResource->GetDesc();
width = desc.Width;
height = desc.Height;
}
const auto validity =
(tag.lifecycle != sl::eOnlyValidNow) ? FG_ResourceValidity::UntilPresent : FG_ResourceValidity::ValidNow;
fgOutput->SetResource(FG_ResourceType::Distortion, cmdBuffer, distortionFieldResource, tag.extent.width,
tag.extent.height, (D3D12_RESOURCE_STATES) tag.resource->state, validity);
fgOutput->SetResource(FG_ResourceType::Distortion, cmdBuffer, distortionFieldResource, width, height,
(D3D12_RESOURCE_STATES) tag.resource->state, validity);
}
return true;
@@ -366,8 +415,7 @@ bool Sl_Inputs_Dx12::dispatchFG()
// Streamline is not 100% clear on if we should multiply by resolution or not.
// But UE games and Dead Rising expect that multiplication to be done, even if the scale is 1.0.
// bool multiplyByResolution = dataCopy.mvecScale.x != 1.f || dataCopy.mvecScale.y
// != 1.f;
// bool multiplyByResolution = dataCopy.mvecScale.x != 1.f || dataCopy.mvecScale.y != 1.f;
bool multiplyByResolution = true;
if (multiplyByResolution)
{
OptiScaler/framegen/Streamline_Inputs_Dx12.h
+1 -1
View File
@@ -12,6 +12,7 @@ class Sl_Inputs_Dx12
std::optional<sl::Constants> slConstants[BUFFER_COUNT] {};
sl::EngineType engineType = sl::EngineType::eCount;
// TODO: make "sent" map/array
bool depthSent = false;
bool hudlessSent = false;
bool mvsSent = false;
@@ -21,7 +22,6 @@ class Sl_Inputs_Dx12
bool distortionFieldRequired = false;
bool dispatched = false;
bool setConstantsSameFrameId = false;
bool frameBasedTracking = false;
uint32_t indexToFrameIdMapping[BUFFER_COUNT] {};
OptiScaler/framegen/xefg/XeFG_Dx12.cpp
+7 -1
View File
@@ -658,10 +658,16 @@ void XeFG_Dx12::SetResource(FG_ResourceType type, ID3D12GraphicsCommandList* cmd
xefg_swapchain_d3d12_resource_data_t resourceParam = GetResourceData(type);
// HACK: XeFG docs lie and cmd list is technically required as it checks for it
// But it doesn't seem to use it when the validity is UNTIL_NEXT_PRESENT
if (cmdList == nullptr && resourceParam.validity == XEFG_SWAPCHAIN_RV_UNTIL_NEXT_PRESENT)
cmdList = (ID3D12GraphicsCommandList*) 1;
auto result = XeFGProxy::D3D12TagFrameResource()(_swapChainContext, cmdList, _frameCount, &resourceParam);
if (result != XEFG_SWAPCHAIN_RESULT_SUCCESS)
{
LOG_ERROR("D3D12TagFrameResource Depth error: {} ({})", magic_enum::enum_name(result), (UINT) result);
LOG_ERROR("D3D12TagFrameResource {} error: {} ({})", magic_enum::enum_name(type), magic_enum::enum_name(result),
(UINT) result);
return;
}
OptiScaler/hooks/HooksDx.cpp
+7 -14
View File
@@ -217,8 +217,6 @@ static HRESULT hkFGPresent(void* This, UINT SyncInterval, UINT Flags)
LOG_DEBUG("_frameCounter: {}, flags: {:X}, Frametime: {}", _frameCounter, Flags, ftDelta);
}
IFGFeature_Dx12* fg = State::Instance().currentFG;
if (willPresent)
{
if (State::Instance().activeFgInput == FGInput::Upscaler && HooksDx::dx12UpscaleTrig &&
@@ -257,11 +255,14 @@ static HRESULT hkFGPresent(void* This, UINT SyncInterval, UINT Flags)
HooksDx::readbackBuffer->Unmap(0, nullptr);
HooksDx::dx12UpscaleTrig = false;
}
if (State::Instance().slFGInputs.readyForDispatch() && !State::Instance().slFGInputs.dispatchFG())
LOG_DEBUG("Streamline FG was not dispatched");
State::Instance().slFGInputs.markLastSendAsRequired();
}
if (State::Instance().slFGInputs.readyForDispatch())
State::Instance().slFGInputs.dispatchFG();
IFGFeature_Dx12* fg = State::Instance().currentFG;
bool mutexUsed = false;
if (willPresent && fg != nullptr && fg->IsActive() &&
Config::Instance()->FGUseMutexForSwapchain.value_or_default() && fg->Mutex.getOwner() != 2)
@@ -277,14 +278,6 @@ static HRESULT hkFGPresent(void* This, UINT SyncInterval, UINT Flags)
fg->Present();
}
State::Instance().slFGInputs.markLastSendAsRequired();
// if (willPresent && State::Instance().activeFgInput == FGInput::DLSSG &&
// State::Instance().slFGInputs.readyForDispatch() && fg != nullptr)
//{
// State::Instance().slFGInputs.dispatchFG(nullptr);
// }
if (willPresent)
{
ResTrack_Dx12::ClearPossibleHudless();
@@ -310,7 +303,7 @@ static HRESULT hkFGPresent(void* This, UINT SyncInterval, UINT Flags)
Hudfix_Dx12::PresentEnd();
if (!State::Instance().reflexLimitsFps && State::Instance().activeFgOutput != FGOutput::NoFG)
if (willPresent && !State::Instance().reflexLimitsFps && State::Instance().activeFgOutput != FGOutput::NoFG)
FrameLimit::sleep(fg != nullptr ? fg->IsActive() : false);
if (mutexUsed && fg != nullptr)
OptiScaler/hooks/Kernel_Hooks.h
+1 -1
View File
@@ -621,7 +621,7 @@ class KernelHooks
// Try to catch something like this:
// C:\ProgramData/NVIDIA/NGX/models/sl_dlss_0/versions/133120/files/190_E658703.dll
if (CheckDllNameW(&lcaseLibName, &slDlssNamesW) ||
(lcaseLibName.contains(L"/versions/") && lcaseLibName.contains(L"/sl_dlss_")))
(lcaseLibName.contains(L"/versions/") && lcaseLibName.contains(L"/sl_dlss_0")))
{
auto dlssModule = KernelBaseProxy::LoadLibraryExW_()(lpLibFullPath, NULL, 0);
OptiScaler/hooks/Streamline_Hooks.cpp
+21 -4
View File
@@ -13,6 +13,7 @@
#include "include/sl.param/parameters.h"
sl::RenderAPI StreamlineHooks::renderApi = sl::RenderAPI::eCount;
std::mutex StreamlineHooks::setConstantsMutex {};
// interposer
decltype(&slInit) StreamlineHooks::o_slInit = nullptr;
@@ -128,7 +129,7 @@ sl::Result StreamlineHooks::hkslSetTag(sl::ViewportHandle& viewport, sl::Resourc
for (uint32_t i = 0; i < numTags; i++)
{
if (tags[i].resource == nullptr)
if (tags[i].resource == nullptr || tags[i].resource->native == nullptr)
{
LOG_TRACE("Resource of type: {} is null, continuing", tags[i].type);
continue;
@@ -414,9 +415,8 @@ bool StreamlineHooks::hkdlssg_slOnPluginLoad(void* params, const char* loaderJSO
sl::Result StreamlineHooks::hkslSetConstants(const sl::Constants& values, const sl::FrameToken& frame,
const sl::ViewportHandle& viewport)
{
unsigned int frameIndex = frame;
LOG_TRACE("called with frameIndex: {}", frameIndex);
std::scoped_lock lock(setConstantsMutex);
LOG_TRACE("called with frameIndex: {}, viewport: {}", (unsigned int) frame, (unsigned int) viewport);
State::Instance().slFGInputs.setConstants(values, (uint32_t) frame);
@@ -426,8 +426,25 @@ sl::Result StreamlineHooks::hkslSetConstants(const sl::Constants& values, const
bool StreamlineHooks::hkcommon_slOnPluginLoad(void* params, const char* loaderJSON, const char** pluginJSON)
{
LOG_FUNC();
// TODO: do it better than "static" and hoping for the best
static std::string config;
auto result = o_common_slOnPluginLoad(params, loaderJSON, pluginJSON);
// Completely disables Streamline hooks
// if (true)
//{
// nlohmann::json configJson = nlohmann::json::parse(*pluginJSON);
// configJson["hooks"].clear();
// configJson["exclusive_hooks"].clear();
// config = configJson.dump();
// *pluginJSON = config.c_str();
//}
if (Config::Instance()->StreamlineSpoofing.value_or_default())
setSystemCapsArch((sl::param::IParameters*) params, UINT_MAX);
OptiScaler/hooks/Streamline_Hooks.h
+1
View File
@@ -38,6 +38,7 @@ class StreamlineHooks
private:
static sl::RenderAPI renderApi;
static std::mutex setConstantsMutex;
// Interposer
static decltype(&slInit) o_slInit;
OptiScaler/resource_tracking/ResTrack_dx12.cpp
+1 -1
View File
@@ -1820,7 +1820,7 @@ void ResTrack_Dx12::HookDevice(ID3D12Device* device)
o_CopyDescriptorsSimple = (PFN_CopyDescriptorsSimple) pVTable[24];
// Apply the detour
if (o_CreateDescriptorHeap != nullptr)
if (o_CreateDescriptorHeap != nullptr && State::Instance().activeFgInput == FGInput::Upscaler)
{
DetourTransactionBegin();
DetourUpdateThread(GetCurrentThread());
OptiScaler/shaders/hudless_compare/HC_Dx12.cpp
+1 -1
View File
@@ -318,7 +318,7 @@ bool HC_Dx12::Dispatch(IDXGISwapChain3* sc, ID3D12CommandQueue* queue, ID3D12Res
// Check Hudless Buffer
D3D12_RESOURCE_DESC hudlessDesc = hudless->GetDesc();
if (hudlessDesc.Format != scDesc.BufferDesc.Format || hudlessDesc.Width != scDesc.BufferDesc.Width ||
if (/*hudlessDesc.Format != scDesc.BufferDesc.Format ||*/ hudlessDesc.Width != scDesc.BufferDesc.Width ||
hudlessDesc.Height != scDesc.BufferDesc.Height)
{
return false;